Crate unimock

source ·
Expand description

unimock is a library for defining mock implementations of traits.

Mocking, in a broad sense, is a way to control API behaviour during test execution.

The uni in unimock indicates one-ness: All mockable traits are implemented by a single type, Unimock. This design allows for a great flexibility in coding style, as will be demonstrated further down.

The first code example is the smallest possible use of unimock:

use unimock::*;

trait Foo {}

fn takes_foo(foo: impl Foo) {}

  1. trait Foo is declared with the #[unimock] attribute which makes its behaviour mockable.
  2. fn takes_foo accepts some type that implements the trait. This function adheres to zero-cost Inversion of Control/Dependency Inversion.
  3. A mock instantiation by calling Unimock::new(()), which crates a Unimock value which is passed into takes_foo.

The new function takes an argument called setup (implementing Clause), in this case the unit value (). The setup argument is what behaviour is being mocked, in this case nothing at all. Foo contains no methods, so there is no behaviour to mock.

§Methods and behaviour mocking

In order to be somewhat useful, the traits we abstract over should contain some methods. In a unit test for some function, we’d like to mock the behaviour of that function’s dependencies (expressed as trait bounds).

Given some trait,

trait Foo {
    fn foo(&self) -> i32;

we would like to tell unimock what Foo::foo’s behaviour will be, i.e. what it will return. In order to do that, we first need to refer to the method. In Rust, trait methods aren’t reified entities, they are not types nor values, so they cannot be referred to in code. We need to tell unimock to expose a separate mocking API. This API will be created in form of a new module, which is named by passing e.g. api=TraitMock to the unimock macro invocation.

Each of the trait’s original methods will get exported as mock config entrypoints through this module: For example TraitMock::method. method is a type that will implement MockFn, which is the entrypoint for creating a Clause:

trait Foo {
    fn foo(&self) -> i32;

fn test_me(foo: impl Foo) -> i32 {

let clause = FooMock::foo.each_call(matching!()).returns(1337);

assert_eq!(1337, test_me(Unimock::new(clause)));

Clause construction is a type-state machine that in this example goes through two steps:

  1. FooMock::foo.each_call(matching!()): Define a call pattern. Each call to Foo::foo that matches the empty argument list (i.e. always matching, since the method is parameter-less).
  2. .returns(1337): Each matching call will return the value 1337.

In this example there is only one clause.

§Call patterns (matching inputs)

It is common to want to control how a function will respond in relation to what input is given to it! Inputs are matched by a function that receives the inputs as a tuple, and returns whether it matched as a bool. A specific MockFn together with an input matcher is referred to as a call pattern from now on.

The matching! macro provides syntax sugar for argument matching. It has a syntax inspired by the std::matches macro.

Inputs being matched is a condition that needs to be fulfilled in order for the rest of the call pattern to be evaluated.

§Specifying outputs (responses)

Specifying outputs can be done in several ways. The simplest one is returns(some_value). Different ways of specifying outputs are found in build::DefineResponse.

There are different constraints acting on return values based on how the clause gets initialized:

§Mutating inputs

Many traits uses the argument mutation pattern, where there are one or more &mut parameters.

To access the &mut parameters (and mutate them), a function is applied to the call pattern using answers:

let mocked = Unimock::new(
        .answers(&|_, f| write!(f, "mutation!"))

assert_eq!("mutation!", format!("{mocked}"));

The argument to answers is a function with the same signature as the method it mocks, including the self parameter.

§Combining setup clauses

Unimock::new() accepts as argument anything that implements Clause. Basic setup clauses can be combined into composite clauses by using tuples:

trait Foo {
    fn foo(&self, arg: i32) -> i32;

trait Bar {
    fn bar(&self, arg: i32) -> i32;

fn test_me(deps: &(impl Foo + Bar), arg: i32) -> i32 {

                .answers(&|_, arg| arg * 3),
                .some_call(matching!((arg) if *arg > 20))
                .answers(&|_, arg| arg * 2),

// alternatively, define _stubs_ for each method.
// This is a nice way to group methods by introducing a closure scope:
            FooMock::foo.stub(|each| {
      !(_)).answers(&|_, arg| arg * 3);
            BarMock::bar.stub(|each| {
      !((arg) if *arg > 20)).answers(&|_, arg| arg * 2);

In both these examples, the order in which the clauses are specified do not matter, except for input matching. In order for unimock to find the correct response, call patterns will be matched in the sequence they were defined.

§Interaction verifications

Unimock performs interaction verifications using a declarative approach. Expected interactions are configured at construction time, using Clauses. Rust makes it possible to automatically verify things because of RAII and the drop method, which Unimock implements. When a Unimock instance goes out of scope, Rust automatically runs its verification rules.

One verification is always enabled in unimock:

Each MockFn mentioned in some setup clause must be interacted with at least once.

If this requirement is not met, Unimock will panic inside its Drop implementation. The reason is to help avoiding “bit rot” accumulating over time inside test code. When refactoring release code, tests should always follow along and not be overly generic.

In general, clauses do not only encode what behaviour is allowed to happen, but also that this behaviour necessarily must happen.

§Optional call count expectations in call patterns

To make a call count expectation for a specific call pattern, look at Quantify or QuantifyReturnValue, which have methods like once(), n_times(n) and at_least_times(n).

With exact quantification in place, output sequence verifications can be constructed by chaining combinators:!(_)).returns(1).n_times(2).then().returns(2);

The output sequence will be [1, 1, 2, 2, 2, ..]. A call pattern like this must be matched at least 3 times. 2 times because of the first exact output sequence, then at least one time because of the .then() combinator.

§Verifying exact sequence of calls

Exact call sequences may be expressed using strictly ordered clauses. Use next_call to define this kind of call pattern.


All clauses constructed by next_call are expected to be evaluated in the exact sequence they appear in the clause tuple.

Order-sensitive clauses and order-insensitive clauses (like some_call) do not interfere with each other. However, these kinds of clauses cannot be combined for the same MockFn in a single Unimock value.

§Application architecture

Writing larger, testable applications with unimock requires some degree of architectural discipline. We already know how to specify dependencies using trait bounds. But would this scale in practice when several layers are involved? One of the main features of unimock is that all traits are implemented by Unimock. This means that trait bounds can be composed, and we can use one value that implements all our dependencies:

fn some_function(deps: &(impl A + B + C), arg: i32) {
    // ..

In a way, this function resembles a self-receiving function. The deps argument is how the function abstracts over its dependencies. Let’s keep this call convention and let it scale a bit by introducing two layers:

use std::any::Any;

trait A {
    fn a(&self, arg: i32) -> i32;

trait B {
    fn b(&self, arg: i32) -> i32;

fn a(deps: &impl B, arg: i32) -> i32 {
    deps.b(arg) + 1

fn b(deps: &impl Any, arg: i32) -> i32 {
    arg + 1

The dependency from fn a to fn b is completely abstracted away, and in test mode the deps: &impl X gets substituted with deps: &Unimock. But Unimock is only concerned with the testing side of the picture. The previous code snippet is at the extreme end of the loosely-coupled scale: No coupling at all! It shows that unimock is merely a piece in a larger picture. To wire all of this together into a full-fledged runtime solution, without too much boilerplate, reach for the entrait pattern.

§Gated mock implementation

If the trait definition, the uses of the trait bound and the tests all live within the same crate, it’s possible to gate the macro invocation:

#[cfg_attr(test, unimock(api = FooMock))]
trait Foo {}

§Combining release code and mocks: Partial mocks

Unimock can be used to create arbitrarily deep integration tests, mocking away layers only indirectly used. For that to work, unimock needs to know how to call the “real” implementation of traits.

See the documentation of new_partial to see how this works.

Although this can be implemented with unimock directly, it works best with a higher-level macro like entrait.


Unimock can be used in a no_std environment. The "std" feature is enabled by default, and can be removed to enable no_std.

The no_std environment depends on alloc and requires a global allocator. Some unimock features rely on a working implementation of Mutex, and the spin-lock feature enables this for no_std. The critical-section feature is also required for no_std. These two features will likely merge into one in some future breaking release.

§Mock APIs for central crates

Unimock works well when the trait being abstracted over is defined in the same code base as the once that contains the test. The Rust Orphan Rule ensures that a Unimock user cannot define a mock implementation for a trait that is upstream to their project.

For this reason, Unimock has started to move in a direction where it itself defines mock APIs for central crates.

These mock APIs can be found in mock.


§What kinds of things can be mocked with unimock?
  • Traits with any number of methods
  • Traits with generic parameters, although these cannot be lifetime constrained (i.e. need to satisfy T: 'static).
  • Traits with associated types and constants, using #[unimock(type T = Foo; const FOO: T = value;)] syntax.
  • Methods with any self receiver (self, &self, &mut self or arbitrary (e.g. self: Rc<Self>)).
  • Methods that take reference inputs.
  • Methods returning values borrowed from self.
  • Methods returning references to arguments.
  • Methods returning Option<&T>, Result<&T, E> or Vec<&T> for any T that is borrowed from self.
  • Methods returning any tuple combination of self-borrowed or owned elements up to 4 elements.
  • Methods returning a type containing lifetime parameters. For a mocked return they will have to be 'static.
  • Generic methods using either explicit generic params or argument-position impl Trait.
  • Methods that are async or return impl Future.
  • async_trait-annotated traits.
§What kinds of traits or methods cannot be mocked?
  • Static methods, i.e. no self receiver. Static methods with a default body are accepted though, but not mockable.
§Selecting a name for the mock api

Due to macro hygiene, unimock tries to avoid autogenerating any new identifiers that might accidentally create undesired namespace collisions. To avoid user confusion through conjuring up new identifier names out of thin air, the name of the mocking API therefore has to be user-supplied. Although the user is free to choose any name, unimock suggests following a naming convention.

The entity being mocked is a trait, but the mocking API is a module. This introduces a conflict in naming convention style, since traits use CamelCase but modules use snake_case.

The suggested naming convention is using the name of the trait (e.g. Trait) postfixed with Mock: The resulting module should be called TraitMock.

This will make it easier to discover the API, as it shares a common prefix with the name of the trait.

§Methods with default implementations

Methods with default implementations use delegation by default. This means that if a default-implementation-method gets called without having been mentioned in a clause, unimock delegates to its default implementation instead of inducing a panic. Quite often, a typical default implementation will itself delegate back to a required method.

This means that you have control over which part of the trait API you want to mock, the high level or the low level part.

§Associated types

Associated types in traits may be specified using the type keyword in the unimock macro:

#[unimock(api = TraitMock, type A = i32; type B = String;)]
trait Trait {
    type A;
    type B;

Working with associated types in a mock environment like Unimock has its limitations. The nature of associated types is that there is one type per implementation, and there is only one mock implementation, so the type must be chosen carefully.

§Associated constants

Associated constants in traits may be specified using the const keyword in the unimock macro:

#[unimock(api = TraitMock, const FOO: i32 = 42;)]
trait Trait {
    const FOO: i32;

Just like with associated types in Unimock, associated constants have the limitation where there is one value of the const per implementation, and there is only one mock implementation, so the value must be chosen carefully.

§Project goals

§Use only safe Rust

Unimock respects the memory safety and soundness provided by Rust. Sometimes this fact can lead to less than optimal ergonomics.

For example, in order to use .returns(value), the value must (generally) implement Clone, Send, Sync and 'static. If it’s not all of those things, the slightly longer .answers(&|_| value) can be used instead.

§Keep the amount of generated code to a minimum

The unimock API is mainly built around generics and traits, instead of being macro-generated. Any mocking library will likely always require some degree of introspective metaprogramming (like macros), but doing too much of that is likely to become more confusing to users, as well as taking longer to compile. The #[unimock] macro does the minimal things to fill out a few simple trait impls, and that’s it. There are no complex functions or structs that need to be generated.

There is a downside to this approach, though. Rust generics aren’t infinitely flexible, so sometimes it’s possible to misconfigure a mock in a way that the type system is unable to catch up front, resulting in runtime (or rather, test-time) failures.

All things considered, this tradedoff seems sound, because this is only testing, after all.

§Use nice, readable APIs

Unimock’s mocking API has been designed to read like natural english sentences.

This was a fun design challenge, but it arguably also has some real value. It is assumed that code is quicker (and perhaps more fun) to read and write when it resembles real language.


  • Builder pattern types used for defining mocked behaviour.
  • Mock APIs for various crates.
  • Function outputs. Traits for modelling the output of MockFns used with returns.
  • Traits and types used for describing the properties of various mock types.


  • Macro to ease call pattern matching for function arguments. The macro produces a closure reference expression suitable for passing to some_call, etc.


  • A marker type used when Unimock is unable to represent the user’s type.
  • Static information about a method
  • A type whose purpose is to provide mocked behaviour for the traits that it implements.


  • A clause represents a recipe for creating a unimock instance.
  • The main trait used for unimock configuration.

Attribute Macros§

  • Autogenerate mocks for all methods in the annotated traits, and impl it for Unimock.